Abstract:
Banking sector at the core is responsible for holding of financial assets in any economy. Bank Failure has far greater effect on the overall economy of a country than any other businesses. It can quickly spill over to other banks and financial institutions and therefore has a multiplying effect. To avoid such scenarios, rigorous regulations have been put in place along with technology to monitor, track and forecast critical parameters. Various statistical techniques and machine learning approaches have been widely adopted in this context. Banks hire domain experts, who along with their expertise exploit these tools to make decisions and recommend actions to prevent bank failure. Despite such tools and expertise, bank failure has occurred from time to time due to complexity of the problem since it is hard to generalize all the knowledge. Lately, with success of AI across different domains, financial institutions and banks have started to adopt much powerful AI methods to replace old methods. In continuation of the modernization effort, this paper proposes a novel deep recurrent neural network for bank failure prediction. More specifically we propose a four-layer recurrent network with Long Short-Term Memory cells. To validate the proposed algorithm, we collected data of 1139 banks from G7 countries and Australia around global financial crises from 2003 to 2013. In total we have collected 59 ratios and variables over eleven years for each of the bank. The proposed algorithm is compared against baseline implementations of widely adapted SVM and Logistic Regression methods. Empirical results demonstrate the superiority of the proposed approach. The paper concludes with a detailed study of effect and role of different parameters towards bank failure.
